Evaluators and safety organisations · Blog post · Redwood Research
Simulated Control Evaluations
Read at blog.redwoodresearch.org ↗Read at sourceAll rights reserved
Outline · 78 paragraphs
- SCEs are a favorable tradeoff between realism and cost¶10
- 1. Run on a copy of internal deployment¶12
- 2. Edit attacks into recorded transcripts¶16
- Evaluating Control Evaluations¶22
- How to measure qualitative properties?¶27
- Performance Comparisons¶30
- Overview of Empirical Results¶35
- Behavior transfer¶39
- Safety metric transfer¶44
- Sample efficiency¶49
- Lessons from building and evaluating SCEs¶52
- Next steps¶58
- Conclusion¶68